Originally Posted by ikjadoon
So, you're telling me that, for all the PCIe slots in an ATX motherboard, on Skylake, we've only got x20 lanes total? Just off the bat, the GPU takes x16! And, so, the other 5 PCIe slots, they're all fighting over x4?
Man, that's confused me for so long.
Well, thank you for clearing this up. I now understand how SLI works and why "tri-SLI" was always listed as x8 / x8 / x4. For the splitter/expander PCH setup, I guess it's happy to give out as many lanes as it wants, but if you try to maximize your devices, you'll be capped at x20.
You cannot run tri-SLI that way. SLI needs a minimum of x8.
To run tri or quad SLI on a Z170 board, you would need a PLX PEX8747. Granted, with Nvidia no longer supporting 3 and 4 way SLI in games, it is mostly for benchmarkers and they are better off using the HEDT platform. That's especially because high end boards with a PLX tend to be very expensive. I have noticed this generation that even boards that used to such as the Asus Z170 Maximus Extreme and the MSI Z170 XPower no longer have PLX chips. Gigabyte still does though, as does EVGA on the Z170 Classified.
Crossfire does run at x4 though. Even at PCIe 3.0 x4, the penalty isn't that bad.https://www.techpowerup.com/reviews/AMD/R9_Fury_X_PCI-Express_Scaling/
So you could run x8/x8/x4 3-way Crossfire via the PCH. Granted it will be a bit less than x4, depending on what else uses that bandwidth.
Of course, this may not be a huge loss either way. Frame time tests show poor game scaling - save Crysis 3 and perhaps Battlefield 4.http://www.pcper.com/reviews/Graphics-Cards/AMD-Fury-X-vs-NVIDIA-GTX-980-Ti-2-and-3-Way-Multi-GPU-Performance
So in that regard, the CryEngine and Frostbite Engines may be very good for multi-GPU.
This might change dramatically though with DX12, especially if Explicit Multi-Adapter gains steam (and most importantly it is split frame rendering, not alternate frame rendering, so frame times are improved, which is far more important that frame rates.
Originally Posted by Oj010
For short bench sessions yes or a really good one, above average sample for 24/7 use, yes. Average samples hit around 4.5 to 4.6 GHz, some bad samples even as low as 4.4 GHz.
There seems to have been a clockspeed drop compared to Devil's Canyon. Granted, what we don't know is whose overclocks are super stable (ex: could pass Linpak, Prime95 Large FFTs, and the most stringent tests), versus whose are just "gaming stable". Assuming a similar proportion though, it does look like compared to the 4790K, there has been a drop in clock speeds by around 200 MHz, which erases a lot of the gains in IPC of the 6700K.
Top samples on the 4790K hit 4.9-5.0 GHz, but those aren't for everyday and require a delid to pass any serious stability test. Plus you gotta be really luck on the silicon lottery (or buy a binned chip). The era of single digit "sidegrades" has been on us for a while now.
Kind of sad that the DMI 3.0 and the new motherboard features are more exciting than the CPU itself.
Originally Posted by Blameless
However, the CPU to z170 bandwidth is still limited by the DMI 3.0 link, which is essentially PCI-E 3.0 4x and anything that needs to go through the CPU from the PCH has to to go through that link.
If you plug 20GB/s worth of I/O into you PCH and try to use it all at the same time, you'll almost certainly see a huge bottleneck at the DMI, unless the devices attached to the PCH are only talking to each other, and not accessing the CPU, system memory, or anything in the 16x lanes that are directly connected to the CPU.
This is why you can't (or at the very least, shouldn't) run multi-GPU configurations with a GPU plugged into the PCH...you get significantly more latency and only have 4x 3.0 lanes of bandwidth to the CPU (and thus the other card in the system), some of which is already utilized by other devices.
Yes, this is a huge bottleneck. X99 and platforms earlier are even worse off with DMI 2.0. I just wish that X99 had DMI 3.0, but it will have to wait for Skylake E.
Personally I'd argue that even DMI 3.0 is not enough and that more lanes need to be allocated to the PCH. This is especially true considering we now have motherboards that offer features like RAID for 3 M.2 disks. That will quickly bottleneck the DMI 3.0 interface.Edited by CrazyElf - 7/11/16 at 5:30pm